Squeeze-MLLM: A New Breakthrough in Subject-Driven Image Generation Powered by Multimodal Large Language Models

Core Insights: The Squeeze-MLLM framework deeply integrates multimodal large language models (MLLMs) with diffusion models, combining the Dual-Layer Aggregation (DLA) module and a multi-stage denoising strategy. It achieves high-quality text-guided image generation while maintaining subject identity consistency, significantly outperforming existing methods. Basic Information:

• Original author team: Researchers including zsh2000
• Source platform: arXiv
• Publication date: May 25, 2026
• Original paper link: http://arxiv.org/abs/2605.26111v1
• Project homepage: https://zsh2000.github.io/squeeze-mllm-subject-gen/

Subject-driven image generation aims to produce new images that both preserve the subject identity and align with text instructions based on reference images and text prompts. It is applied in scenarios like personalized creation and e-commerce display. Existing methods face two major challenges:

1. Separate encoding limits cross-modal reasoning ability, making it difficult to understand complex semantic relationships between text and images;
2. Lack of effective identity preservation mechanisms, which easily leads to "copy-paste" artifacts. Although some studies have combined MLLMs with diffusion models, they ignore identity preservation, resulting in poor identity consistency in generated results.

Core ideas of the Squeeze-MLLM framework:

• Joint Encoding: Let MLLMs process both text and reference images simultaneously, understanding their relationships in a unified semantic space to avoid mechanical concatenation;
• VAE Identity Condition: Introduce a variational autoencoder to extract fine-grained features, ensuring precise preservation of subject details such as texture and color. The combination of these two achieves an organic unification of semantic understanding and identity preservation.

The Dual-Layer Aggregation (DLA) module aggregates features from different layers of MLLMs:

• Shallow features: Retain details like edges and textures;
• Deep features: Contain semantics such as object categories and scene understanding. By adaptively adjusting weights, it provides rich conditional signals for diffusion models, flexibly adapting to different generation needs (e.g., precise identity preservation or complex semantic understanding).

The multi-stage denoising strategy balances semantics and identity in stages:

• Early stage: Rely on MLLM semantic conditions to establish correct composition and conceptual expression;
• Late stage: Increase the weight of VAE identity conditions to restore subject details. This simulates the human painting process (outlines first, then details), effectively alleviating "copy-paste" artifacts.

Experimental results show that Squeeze-MLLM significantly outperforms existing methods:

• Human preference: Generated images have higher subjective scores, balancing semantic understanding and identity preservation;
• Quantitative metrics: Excellent performance in dimensions like identity preservation, text-image alignment, and image quality;
• Artifact mitigation: Achieves more flexible and natural subject transfer, reducing traces of rigid splicing.

Technical Insights:

1. MLLMs have great potential in visual generation tasks and can handle both text and image inputs uniformly;
2. The multi-level feature fusion strategy of DLA can be applied to multi-scale visual tasks;
3. The dynamic conditional adjustment of multi-stage denoising provides new ideas for conditional control of diffusion models. Future Outlook: Explore more aggregation mechanisms, extend dynamic conditional adjustment to other diffusion applications, and look forward to more intelligent image generation systems.